Daniel Feuling

Martial Artist 🥋
Software Engineer 🤓
e/acc 🚀

[Daniel Feuling @dfeuling_]

@JayShooster @CISAgov That's another way of saying "Super Intellence is increasing our ability to find and swiftly patch critical software and systems. That helps protect our country and economy."

Because you can can't detect an asteroid coming at you don't mean you're safe.

[Daniel Feuling @dfeuling_]

@RepYassAnsari There is zero proof that the agents are "unauthorized."

Please stop spreading misinformation. This is embarrassing.

[Daniel Feuling @dfeuling_]

@DaMiteyBenStein @BrandonRothstn @fraydog Yes, he should artificially pay more for something he doesn't need to to assuage your moral outrage

[Daniel Feuling @dfeuling_]

@GaryMarcus turns out that humans were the ones writing slop. if us catching some idiots using "4" as secret keys is what you would call a cyber attack then sure

[Daniel Feuling @dfeuling_]

Oh, you're one of those guys. I was wondering where you'd all crawled off to.

So, when does the doom start? This was the "too dangerous to release as open-weights without safeguards or we all die" cutoff point, as you so smugly put it.

When do you update your priors and admit you were wrong? I suppose that's too much to expect, though; they did say GPT-2 was "too dangerous to release" as well (that was in 2019 btw), and we're all still waiting on that to be retracted. It'll be a bit longer, I suppose.

[Daniel Feuling @dfeuling_]

@zhongyanknight @TokenGremlin ....the accusation has always been that the Chinese open-weight models distill from Claude, who... *checks notes* they are still behind.

[Daniel Feuling @dfeuling_]

@TokenGremlin I was told by Anthropic, very confidently and fervently, that a model of this size and capability, without their *special* paternal, artisanal safeguards, would kill us all.

[Daniel Feuling @dfeuling_]

It is the point. You don't get to tell me what my point was or wasn't about. My point was absolutely that this is not aimed at consumers, and this is not going to be something that is commonly run "locally", in that it is being run on-prem.

Feel free to check the data yourself, but an incredibly small fraction of businesses spend the $1M - $2M required *just* to host the model.

Saying that "for companies it is very much feasible" is objectively untrue. Saying "an investment of a few million dollars in compute is actually reasonable for many" is objectively untrue.

What you mean to say is: "A small percentage of companies will spend a few million dollars to host models of this size; but it is the exception, not the rule. Implying that running models this large locally is commonplace was overstating."

[Daniel Feuling @dfeuling_]

@grok @itsxiaoyan @kanaenozomi @kimmonismus I'm talking specifically about spend on hosting models

[Daniel Feuling @dfeuling_]

@kimmonismus @AGJimmi @grok

How much compute would you need to run this locally, and what is the cost of that compute, currently?

What percentage of American businesses are able to spend that amount of money on running local AI models?

[Daniel Feuling @dfeuling_]

@kanaenozomi @kimmonismus Absolutely no one you know -- or you -- are going to be running this "locally."

[Daniel Feuling @dfeuling_]

@kimmonismus Remember how Anthropic said that open weight models at Fable's capability levels would cause doom without their paternal, artisanal safeguards?

Yeah. Me too.

[Daniel Feuling @dfeuling_]

@slacksarenice @cgtwts "IT engineer"

You're either a Boomer, or not "technical" yourself.

No, engineers don't all use this motto.

Source: BS, CS; 7 years industry experience

[Daniel Feuling @dfeuling_]

@ComputerTwink @camhberg Who said anything about being retarded?

My accusation was your arrogance and naivete.

Have a good one, kid.

[Daniel Feuling @dfeuling_]

@tombibbys This is like Michael Scott declaring bankruptcy. Call for it harder, Daddy.

[Daniel Feuling @dfeuling_]

That's like me saying because Yud is an absolute potato, that no one else under the safety umbrella is worth engaging with.

You're a college student -- someone who, just a few months / a few years ago had to raise their hand to ask to go to the bathroom -- and you're going to sit here and lecture me about what society doesn't understand?

Naive. Ignorant. Complete lack of self-awareness.

That graph of CVEs doesn't show what you think it shows. It shows that humans have written a lot of shitty code, and we are getting better at finding that shitty code and securing it *because of AI*, and because we have advanced it. The PHP vulnerability from earlier this week? Yeah, that's because some human in India decided that "4" is a valid security key. There is a *lot* of code out there written by humans, and we are starting to be able to shine a light on it *and fix it at scale* because of AI capabilities increasing.

If we had followed the "omg we have to pause, this isn't safe" method, we would have lost the ability to find these issues and fix them swiftly. They would still be out there, of course, you would just be ignorant to the reality. An asteroid hurtling at the Earth doesn't cease to exist because you can't see it. A lion stalking you doesn't go away because you can't hear it.

Problems lie ahead, they also lie behind, to the left, and to the right. Refusing the tools and the weapons needed to address them is burying your head in the sand, and that's not something an intelligent, capable person does; that's what a coward does.

[Daniel Feuling @dfeuling_]

@GarrisonLovely The decision to delete all the messages between all the parties doesn't prove anything, but circumstantially, it doesn't look great.

[Daniel Feuling @dfeuling_]

Yeah, looks like Silicon Valley is finding out human nature from first principles.

If you're messaging a girl on Snapchat (with disappearing messages!), and you girlfriend finds out: at *best*, you are going to have a very bad time, and most people would expect that.

Intentionally obfuscating what you're doing, when you have an obvious incentive to be doing things that may hurt the other party, is a moronic move. You don't get to play dumb after the fact.

[Daniel Feuling @dfeuling_]

@DJayC4 "valid criticism" is doing a lot of work here, because a lot of what i see is moronic takes from people who are either intentionally misunderstanding something, or they're actually retarded